[foota]

Seems interesting... how did you land on valgrind vs some other means of simulation? Looking at valgrind, it sounds like cachegrind? Imo seems like the biggest gap is non instruction, non cache sources of latency, like mutex contention or kernel slowness? (Or does it capture kernel delays?)

Pur most recent performance issues have been from someone accidentally creating a new thread pool in a request, from generating tons of stack traces in an error handling path, and from some thread hop delays. Sounds like the first two would probably be caught but maybe not the third?